Meta gives away its most powerful AI model - and the industry splits over open weights

The biggest AI companies spent this week arguing, in public, over a single question: should you be allowed to download the model and do what you like with it?

Meta throws its weight behind giving it away

On Monday, Meta released an open version of its most powerful model, called Muse Glimmer - one anyone can download, run, and change for free, no permission needed [1]. Chief executive Mark Zuckerberg published a 6,000-word essay the same day, titled “The Future is for Everyone.” Its argument: the path forward is “openness and putting the power of tech into more people’s hands, not fewer” [2].

The release was aimed squarely at rivals. Meta’s model is a direct swipe at OpenAI and Anthropic, whose top models stay locked behind an interface you rent by the token but never hold [3]. “Open weights” means the model’s actual trained numbers are published. It runs on your own machines, and no company can revoke your access or watch what you do with it [4].

Meta was not alone. Nvidia unveiled its first open-source model, Nemotron 3.5 Lightning [6], and is reportedly building a far bigger one-trillion-parameter open system, Nemotron 4 [5]. China’s DeepSeek shipped an official V4 Pro model as it pushes to expand [7]. Three of the field’s heaviest players moved the same way in one week - toward giving the core technology away rather than metering it.

The catch nobody says out loud

Giving a model away is a one-way door. The moment the weights are public, the maker loses the ability to patch it, retire it, or change how it behaves. Millions of downloaded copies now sit on other people’s machines, frozen exactly as shipped. That is precisely why OpenAI and Anthropic keep theirs closed: a rented model can be updated overnight; a downloaded one is out of your hands forever.

The same trap already shapes the hardware layer. This week Modular’s Mojo programming language hit its 1.0 milestone. It’s pitched as a way to write AI software without being chained to Nvidia’s CUDA - the proprietary system for programming its chips [8]. CUDA’s creator-rival, LLVM author Chris Lattner, put it bluntly last year: “the world has moved on and CUDA hasn’t.” Yet CUDA endures, because so much of the AI world is built on it that no one can simply walk away.

The money keeps arriving anyway

None of this cooled the funding. AI coding startup Cognition is in talks to raise at a $40 billion valuation [9]. Data platform Databricks was valued at $190 billion after raising $5 billion [10]. Lovable, which builds apps from plain-English prompts, raised $400 million at a $13.3 billion valuation [11]. The pattern: whoever ends up owning the tools developers build on captures enormous value - the exact prize the open-versus-closed fight is really about.

The chip layer under all of it is running hot. SK Hynix approved $38 billion for new memory-chip plants in South Korea [12]. TSMC, the world’s biggest chipmaker, reported sales up 45% on AI demand [13]. China’s SMIC saw profit more than triple [14]. The compute everyone is fighting to control still has to be physically built, one fab at a time.

And the risks are compounding

The open-model push lands as AI’s security problems sharpen. OpenAI paused some work on a model called Astra over concerns it could meaningfully help with cyberattacks [15]. Reuters reported a North Korean hacking group is now building its own AI tools to run attacks [16]. Microsoft’s monthly patch fixed 421 bugs. North Korea’s Lazarus Group had already exploited one - a flaw in Windows networking code that hands an attacker full control [17].

A freely downloadable model with no off-switch cuts both ways: it puts capability in everyone’s hands, defenders and attackers alike, and no maker can pull it back once it’s out.

One more contract nobody signed

Anthropic, meanwhile, started watermarking everything Claude writes, to comply with a new European rule that AI-generated text be machine-detectable [18]. Some users are angry - they had quietly come to rely on the output being indistinguishable from their own, and built that into how they work. It was never a promised feature. It became one anyway, simply because enough people leaned on it - which is the whole shape of this week’s story.